softmax_kernel.cpp 5.3 KB
Newer Older
H
hanbuhe 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#ifdef SOFTMAX_OP

Z
zhangyang 已提交
17 18
#include "operators/kernel/softmax_kernel.h"
#include "operators/kernel/central-arm-func/softmax_arm_func.h"
19

H
hanbuhe 已提交
20 21 22 23
namespace paddle_mobile {
namespace operators {

template <>
N
nhzlx 已提交
24
bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
25 26 27
  auto input = const_cast<LoDTensor *>(param->InputX());
  auto dims = framework::vectorize(input->dims());

28 29 30 31
  auto out = param->Out();
  out->Resize(framework::make_ddim(dims));
  out->mutable_data<int8_t>(framework::make_ddim(dims));
  fpga::format_ofm(out);
32 33 34 35 36 37 38 39 40 41 42

  PADDLE_MOBILE_ENFORCE(input->dims().size() == 4,
                        "Softmax should have 4-order input");

  auto channel = dims[3];
  if (channel == 1) {  // This input is generated by FC op, dims = [N C 1 1]
    PADDLE_MOBILE_ENFORCE(dims[2] == 1, "Softmax input must come from FC op");
    dims[3] = dims[1];
    dims[1] = 1;
  }
  input->Resize(framework::make_ddim(dims));
43 44 45 46
  if ((channel == 2) && (input->type() == type_id<int8_t>())) {
    auto input_ptr = input->data<int8_t>();
    float Si = input->scale[0];
    int16_t slope = fpga::fp32_2_fp16(Si / 127);
47
    fpga::format_ofm(out);
48 49 50 51 52 53 54 55 56
    fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
    args.input_layout_type = fpga::LAYOUT_HWC;
    args.output_layout_type = fpga::LAYOUT_CHW;
    args.input_data_type = fpga::DATA_TYPE_FP16;
    args.output_data_type = fpga::DATA_TYPE_FP16;
    args.image.address = input_ptr;
    args.image.height = (uint32_t)input->dims()[1];
    args.image.width = (uint32_t)input->dims()[2];
    args.image.channels = (uint32_t)input->dims()[3];
57
    args.output.address = out->data<int8_t>();
58 59
    args.output.scale_address = out->scale;
    args.output.activation.activation_type = fpga::SOFTMAX;
60
    args.output.activation.leaky_relu_negative_slope = slope;
61
    param->SetFpgaArgs(args);
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
  } else if (input->type() == type_id<int8_t>()) {
    auto float_input_x = param->float_input_x_;
    float_input_x = std::make_shared<Tensor>();
    float_input_x->Resize(input->dims());
    float_input_x->init(type_id<float>().hash_code());
    fpga::format_ofm(float_input_x.get());
    auto float_out = param->float_out;
    float_out = std::make_shared<Tensor>();
    float_out->Resize(input->dims());
    float_out->init(type_id<float>().hash_code());
    fpga::format_ofm(float_out.get());
  } else {
    auto float_out = param->float_out;
    float_out = std::make_shared<Tensor>();
    float_out->Resize(input->dims());
    float_out->init(type_id<float>().hash_code());
    fpga::format_ofm(float_out.get());
79 80
  }

H
hanbuhe 已提交
81 82 83 84
  return true;
}

template <>
85
void SoftmaxKernel<FPGA, float>::Compute(const SoftmaxParam<FPGA> &param) {
86
  auto *in_x = (param.InputX());
87 88 89 90 91 92
  auto dims = in_x->dims();
  auto n = dims[0];
  auto h = dims[1];
  auto w = dims[2];
  auto c = dims[3];
  if ((c == 2) && (in_x->type() == type_id<int8_t>())) {
93
    fpga::PerformBypass(param.FpgaArgs());
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
  } else if (in_x->type() == type_id<int8_t>()) {
    auto in_data = in_x->data<int8_t>();
    float Si = in_x->scale[0];
    Tensor *out = param.Out();
    out->Resize(
        {in_x->dims()[0], out->dims()[1], out->dims()[2], out->dims()[3]});
    auto out_data = out->data<int8_t>();
    auto float_input_x = param.float_input_x_;
    auto float_input_x_data = float_input_x->data<float>();
    int dataNum = n * h * fpga::align_to_x(w * c, IMAGE_ALIGNMENT);
    for (int i = 0; i < dataNum; i++) {
      float_input_x_data[i] = in_data[i] * Si / 127;
    }
    auto float_out = param.float_out;
    auto float_out_data = float_out->data<float>();
    math::SoftmaxFuntor<CPU, float>()(float_input_x.get(), float_out.get());
    for (int i = 0; i < dataNum; i++) {
      float tmp_out = float_out_data[i] * 127;
      out_data[i] = tmp_out < 0 ? (signed char)(tmp_out - 0.5)
                                : (signed char)(tmp_out + 0.5);
114
    }
115
    fpga::fpga_flush(out_data, dataNum * sizeof(int8_t));
116
  } else {
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
    Tensor *out = param.Out();
    out->Resize(
        {in_x->dims()[0], out->dims()[1], out->dims()[2], out->dims()[3]});
    auto out_data = out->data<int8_t>();
    auto float_out = param.float_out;
    float_out = std::make_shared<Tensor>();
    float_out->Resize(in_x->dims());
    float_out->init(type_id<float>().hash_code());
    fpga::format_ofm(float_out.get());
    auto float_out_data = float_out->data<float>();
    math::SoftmaxFuntor<CPU, float>()(in_x, float_out.get());
    int dataNum = n * h * fpga::align_to_x(w * c, IMAGE_ALIGNMENT);
    for (int i = 0; i < dataNum; i++) {
      float tmp_out = float_out_data[i] * 127;
      out_data[i] = tmp_out < 0 ? (signed char)(tmp_out - 0.5)
                                : (signed char)(tmp_out + 0.5);
133
    }
134
    fpga::fpga_flush(out_data, dataNum * sizeof(int8_t));
135
  }
H
hanbuhe 已提交
136 137 138 139 140 141
}

}  // namespace operators
}  // namespace paddle_mobile

#endif